Integrated Frequency References
Alexander Delke is a PhD student in the Department Integrated Circuit Design. (Co)Promotors are prof.dr.ir. B. Nauta and dr.ir. A.J. Annema from the Faculty of Electrical Engineering, Mathematics and Computer Science.
In today’s world, electronic devices like mobile phones and Internet of Things (IoT) gadgets have become indispensable to many aspects of our daily lives. A critical component in nearly every electronic device is the frequency reference, which generates an accurate and stable frequency for other subblocks under various conditions. Frequency references are essential for communication through wired and wireless connections and for data conversion, like analog-to-digital. The required frequency accuracy is application dependent; for example, ±0.03% (±300 ppm) is necessary for SuperSpeed USB and must be achieved over an operating range, for example, the automotive temperature range from −40 °C to 125 °C.
Quartz crystal oscillators (XOs) are the de facto standard in the industry for frequency references, primarily due to their superior frequency accuracy across process voltage temperature (PVT) variation and lifetime. However, quartz crystal, a piezoelectric resonator, needs a vacuum-sealed package to protect against air damping. As a result, it is a bulky external component with a footprint larger than 1 mm2. The rise in battery-powered devices and the demand for a higher level of integration drive the search for low-power, fully integrated, cost-effective frequency references that match the accuracy of XO in standard CMOS technology. Options that can be integrated into standard CMOS technology include RC-based oscillators, LC-based oscillators, and thermal-diffusivity frequency references. However, each of these alternatives has specific disadvantages that prevent them from matching the performance of XO.
The work described in this thesis aims to develop a fully integrated frequency reference in a commercially available CMOS process as a low-cost XO replacement. After a brief introduction to frequency references, the second chapter of this thesis addresses the challenges of frequency references in standard CMOS, discussing sources of short- and long-term frequency variation, such as PVT variation and degradation over time. Furthermore, existing fully integrated frequency reference methods that have the potential to replace XOs are reviewed.
A key challenge for integrated frequency references is the high sensitivity of components like resistors and capacitors to process and temperature variations, e.g., >1000 ppm/°C for front-end resistors. This sensitivity makes it difficult to achieve the same level of frequency accuracy as XOs, which typically maintain accuracy within ±100 ppm. While active tuning can reduce temperature-induced frequency drift, both the tuning circuits and temperature coefficient of the frequency-determining components are subject to process variations, often requiring multiple post-production temperature trimming steps to achieve the required accuracy. These multiple trims add significant cost, highlighting the need for a frequency reference with minimal process sensitivity in its temperature coefficient, ideally requiring only a single temperature trim (1T-trim) to maintain low production costs.
Chapter 3 of this thesis analyzes and compares the temperature coefficient and its process variation of the cross-coupled LC oscillator and Colpitts LC oscillator. The analyses show that the temperature coefficient of the Colpitts LC oscillator is ten times less sensitive to process variation than that of a cross-coupled LC oscillator. This can be leveraged to generate a well-behaved frequency suitable for a single-trim frequency reference. A prototype in a 130 nm high-voltage (HV) CMOS SOI process achieves a maximum of ±70 ppm frequency inaccuracy over a temperature range from −50 °C to 170 °C after a sample-specific 1T-trim with batch-calibration over 48 samples from three wafers (one batch).
Temperature sensors are part of many frequency references to minimize the temperature-induced frequency drift. Consequently, inaccuracies in these sensors can significantly impact the overall frequency error budget. Chapter 4 of this thesis focuses on designing a temperature sensor with an accuracy better than ±0.5 °C, without adding excessive power consumption, chip area, or extra temperature trimming points w.r.t. the frequency reference presented in Chapter 5. In this design, the temperature reading is based on two temperature-dependent currents generated in a BJT-based sensing front end employing dynamic error cancellation techniques to improve accuracy, digitized by a dual-slope ADC. A prototype in a 130 nm HV CMOS SOI process achieves an inaccuracy of ±0.3 °C (3σ) from −40 °C to 130 °C after a sample-specific 1T-trim and batch calibration.
A Colpitts LC oscillator can match the frequency accuracy of low-cost XOs across a wide range of operating conditions. However, it consumes significantly more power and typically, in a standard CMOS process, operates at frequencies higher than those of standard XOs. To serve as a viable replacement for XO, the frequency accuracy provided by the Colpitts LC oscillator must be achieved at lower frequencies and power consumption. Chapter 5 of the thesis describes a frequency reference system design that utilizes a production-trimmed Colpitts LC oscillator and an in-field periodically calibrated second, more low-power oscillator. This system addresses the high power consumption issue of the Colpitts LC oscillator while preserving its high accuracy/stability but at a lower frequency. The presented prototype achieves a maximum frequency inaccuracy of ±93 ppm over a temperature range from −63 °C to 165 °C while consuming 210 μW from a single 3.3 V supply at an output frequency of 70 MHz.
In conclusion, the techniques presented in this thesis show that frequency references in a standard CMOS process can match XOs in terms of frequency accuracy without excessive post-production trimming and power consumption.
